Choose hardware for running AI on your own machine

Macs, DGX Spark-class boxes, 128 GB mini PCs and RTX desktops, compared on the numbers that decide what a model feels like: whether it fits, how long the first word takes, and how fast the answer comes. Every number is labelled, and our own measurements sit next to what the spec sheets promise.

• Is your current Mac enough? Pick your chip and memory; see which models fit and how fast they run. Many readers need nothing new.
• Which machine for which person The one number that decides it for each class, with current prices.
• Compare two machines Side by side: memory, bandwidth, answer speed and prompt time.

What we measured

38 models on three Macs we own. Dense models reach about 78% of the speed their memory bandwidth allows; mixture-of-experts models about 41%. Spec sheets cannot tell you that. See the measurements.

How to read the numbers

• spec published by the maker, linked.
• cited a third party's measurement or report, linked and dated.
• derived computed from specs with a rule we show; speeds are upper bounds, times lower bounds.
• measured measured by us, on the Macs named on the Measured page.
• assumption a working assumption, stated so you can change it.

Reference: the spec table · from spec sheet to meaning · how we test and how we earn